CyberGuard dictionary / AI Security
Model Exfiltration
Model exfiltration is unauthorized removal of model files, weights, parameters, or enough model behavior to reproduce a valuable system.
In simple terms
An attacker may steal the actual model artifact or repeatedly query an exposed model to build an imitation. These are different paths with similar ownership and security consequences.
Technical explanation
Controls include protecting model storage and deployment credentials, limiting query access, monitoring unusual extraction patterns, and separating development artifacts from production interfaces. Rate limits alone do not stop every theft path.
Theft can involve direct artifact access, compromised storage, or extensive probing of an exposed interface. Controls should match the route, including credentials, egress, query monitoring, and artifact permissions.
Examples in practice
A compromised build account downloads a private model checkpoint from a storage bucket that ordinary application users cannot access.
A build credential unexpectedly reads a private checkpoint and transfers it to an outside location. Audit logs identify the access path, and the team revokes the credential and checks other artifacts.
Why it matters
Models can contain substantial proprietary value and may also expose information about their training or operation.
VERIFY & LEARN MORE
Sources and further reading
Published 2026-10-15 · Last updated 2026-10-15 · CyberGuard editorial content. Source links provide context; they do not imply endorsement.